Unboxing the Black Box of Attention Mechanisms in Remote Sensing Big Data Using XAI

Abstract

This paper presents exploratory work looking into the effectiveness of attention mechanisms (AMs) in improving the task of building segmentation based on convolutional neural network (CNN) backbones. Firstly, we evaluate the effectiveness of CNN-based architectures with and without AMs. Secondly, we attempt to interpret the results produced by the CNNs using explainable artificial intelligence (XAI) methods. We compare CNNs with and without (vanilla) AMs for buildings detection. Five metrics are calculated, namely F1-score, precision, recall, intersection over union (IoU) and overall accuracy (OA). For the XAI portion of this work, the methods of Layer Gradient X activation and Layer DeepLIFT are used to explore the internal AMs and their overall effects on the network. Qualitative evaluation is based on color-coded value attribution to assess how the AMs facilitate the CNNs in performing buildings classification. We look at the effects of employing five AM algorithms, namely (i) squeeze and excitation (SE), (ii) convolution attention block module (CBAM), (iii) triplet attention, (iv) shuffle attention (SA), and (v) efficient channel attention (ECA). Experimental results indicate that AMs generally and markedly improve the quantitative metrics, with the attribution visualization results of XAI methods agreeing with the quantitative metrics.